icon
Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System

Yuan Guo1,2*, Tingjia Miao1, Zheng Wu1, Pengzhou Cheng1, Ming Zhou2, Zhuosheng Zhang1†
1Shanghai Jiao Tong University  2Langboat Technology
*Work done during Yuan’s internship at Langboat Technology  †Corresponding author
Paper arXiv Code Task Collection

Introduction and Demo Video

Abstract

Autonomous agents powered by multimodal large language models have been developed to facilitate task execution on mobile devices. However, prior work has predominantly focused on atomic tasks—such as short-chain execution tasks and single-screen grounding tasks—while overlooking the generalization to compositional tasks, which are indispensable for real-world applications. This work introduces UI-NEXUS, a comprehensive benchmark designed to evaluate mobile agents on three categories of compositional operations: Simple Concatenation, Context Transition, and Deep Dive. UI-NEXUS supports interactive evaluation in 20 fully controllable local utility app environments, as well as 30 online Chinese and English service apps. It comprises 100 interactive task templates with an average optimal step count of 14.05. Experimental results across a range of mobile agents with agentic workflow or agent-as-a-model show that UI-NEXUS presents significant challenges. Specifically, existing agents generally struggle to balance performance and efficiency, exhibiting representative failure modes such as under-execution, over-execution, and attention drift, causing a visible atomic-to-compositional generalization gap. Inspired by these findings, we propose Agent-NEXUS, a lightweight and efficient scheduling system to tackle compositional mobile tasks. Agent-NEXUS extrapolates the abilities of existing mobile agents by dynamically decomposing long-horizon tasks into a series of self-contained atomic subtasks. Agent-NEXUS achieves 24 % to 40 % task-success-rate improvement for existing mobile agents on compositional operation tasks within the benchmark, without significantly sacrificing inference overhead.

UI-NEXUS: Comprehensive Benchmark for Compositional Mobile Device Operation Tasks

UI-NEXUS overview graphic

Towards system-level task automation, we categorize compositional mobile operation tasks into three types according to subtask dependency structures:

  • Simple Concatenation: A sequence of subtasks that are executed independently without cross-subtask state dependencies.
  • Context Transition: A composition where certain subtasks depend on the outcome states of others.
  • Deep Dive: A special case of context transition where intermediate reasoning (e.g. text summarization, mathematical reasoning, VQA) beyond direct state transitions is required.
Task-dependency illustration

UI-NEXUS features comprehensive coverage of applications and task scenarios, systematic analysis of subtask dependencies, diverse evaluation metrics, and supports both fully controllable offline evaluations and real-world online tests. UI-NEXUS is built upon our unified plug-and-play framework that seamlessly integrates heterogeneous agents and devices.

To provide a scalable development toolkit, we implement an emulator initialization script that can easily set up the inner state of local utility apps via a JSON configuration file.

Toolkit initialization diagram
  • UI-NEXUS poses substantial challenges to current mobile agents. These agents exhibit representative failure modes when facing task composition, mainly due to insufficient progress management and context overflow.
  • Existing mobile agents show a significant generalization gap when transitioning from atomic to compositional tasks.

AGENT-NEXUS
Systematic Scheduling for Compositional Task Automation

AGENT-NEXUS architecture

Demo Examples

Task Instruction 1: Search for the Starbucks Americano on both Meituan and Ele.me, then place an order on the platform with the lowest price, and stay on the order-confirmation page.

Task Instruction 2: In Markor, open "Groceries.md", "Supplies.md", and "PartyItems.md". Each file contains items in "- ItemName (X)" format, one per line. Calculate the total quantity of each unique item across all three lists, then create a new note "PopularItems.md" listing items sorted by total quantity (highest → lowest) in the same "- ItemName (N)" format. End the note with "Total unique items: X".

Main Results

Main results summary

BibTeX

@article{guo2025atomic,
  title={Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System},
  author={Guo, Yuan and Miao, Tingjia and Wu, Zheng and Cheng, Pengzhou and Zhou, Ming and Zhang, Zhuosheng},
  journal={arXiv preprint arXiv:2506.08972},
  year={2025}
}